microsoft--agent-framework/python/samples/04-hosting/durabletask/11_subworkflow

Composed Workflow (Sub-Workflow) on a Standalone Durable Task Worker

This sample demonstrates workflow composition on a standalone Durable Task worker: an inner agent-framework Workflow is embedded as a node inside an outer Workflow using WorkflowExecutor. On the durable host, the inner workflow runs as its own durable child orchestration.

Key Concepts Demonstrated

• Embedding one Workflow inside another with WorkflowExecutor(inner_workflow, id=...).
• A single DurableAIAgentWorker.configure_workflow(outer_workflow) call walks the composition and auto-registers a durable orchestration for each workflow:
  • dafx-review_pipeline — the outer workflow.
  • dafx-sentiment_analysis — the inner workflow, run as a durable child orchestration when the outer workflow reaches the WorkflowExecutor node.
• Per-workflow scoping: each workflow's agent executors become durable entities and its non-agent executors become durable activities, named per workflow so the same executor id in two workflows never collides.
• Output forwarding: the inner workflow yields a string and, because allow_direct_output is left at its default (False), that output is forwarded to the outer workflow as a message delivered to the reporter executor.

Composition Layout

review_pipeline (outer)
  intake (executor)
    -> sentiment_sub = WorkflowExecutor(sentiment_analysis)
         sentiment_agent (agent) -> sentiment_formatter (executor)
    -> reporter (executor)

Environment Setup

See the README.md in the parent directory for environment setup.

This sample uses Azure AI Foundry credentials:

• FOUNDRY_PROJECT_ENDPOINT
• FOUNDRY_MODEL

It also needs a Durable Task Scheduler. For local development, start the emulator (defaults to http://localhost:8080):

docker run -d -p 8080:8080 -p 8082:8082 mcr.microsoft.com/dts/dts-emulator:latest

Running the Sample

Start the worker in one terminal:

cd samples/04-hosting/durabletask/11_subworkflow
python worker.py

In a second terminal, run the client:

python client.py

The client targets only the outer workflow (review_pipeline); the sub-workflow runs automatically as a child orchestration. Each review flows:

intake → sentiment_sub (child orchestration: sentiment_agent → sentiment_formatter) → reporter → "Review analysis complete -> sentiment: ...".